Write and sign the rules
A file for each deployment sets the agent's scope, role and limits. Your team signs it before the agent runs.
AI agents your auditors can approve
Let AI agents do the work, with policy in the path.
Book a walkthroughRuntime policy enforcement, a named approver where judgment is needed, and a sealed record of every action an agent takes.
Gatehousecheckseachagentactionagainstyoursignedpolicybeforeitruns,sendsthehardcallstoonenamedperson,andkeepsasealedrecordofwhathappened.
Policy engine
Your rules, in one signed file.
It sets the scope, role, limits, review and seal for a deployment. The policy engine turns it into the rules checked before each action runs.
Approvals
Hard calls wait for a named approver.
Below the confidence level you set, the run pauses. The request goes to that approver with the evidence. It is never sent to a shared inbox.
Record
Each run is saved to the audit trail.
It shows what the agent saw, cited and decided, what it cost, and who approved it. Nobody writes any of it by hand.
Data portability
Take the record with you.
The trail, policy history and labelled examples export in open formats. You can read them without Gatehouse installed.
Putting an agent to work
Sign the rules first. The agent can't step outside them.
Check each action
The policy engine looks at each action before it runs. If the rules allow it, it goes ahead. If not, it is held for a person or refused.
Read the sealed record
Each action, hold and decision is saved in order, chained by SHA-256. One query returns the full run. Editing any record breaks the chain.
Your team can check the record without taking our word.
Surehand, the firm that built Gatehouse, leads each walkthrough. It needs no access to your systems.We bring manifest v14 and the sealed record of INV-40218, the simulated run used across this site. You bring one live process and your reviewers' questions. The call takes 30 minutes, with no slides.
Book a walkthroughExamples, each labelled
Three cases and one run, step by step.
- W/01
Claims triage at a regional TPA.
Files the agent is unsure about are held for a named claims handler, with the evidence gathered. By policy, only a person can reject a claim.
Composite case
- 14:31:52
The run opens.
RUN-7F2K4 opens under signed manifest v14. The policy check passes.
Simulated run
- W/02
Invoice matching for a distributor.
Matches the agent is sure of post to the ERP as drafts under a spend limit. An amount over the limit, or a changed supplier bank detail, goes to the AP lead by name first.
Composite case
14:31:58The invoice matches.
INV-40218 matches its purchase order and goods receipt. Read, match, draft and schedule all pass.
Simulated run
- W/03
Policy answers with named escalation.
Each answer cites the clause it rests on. When the agent is unsure, the question goes to the support lead by name.
Composite case
- 14:32:00
The payment is held.
The bank detail was first seen today, so rule bank-detail-change holds the payment. The AP lead is asked to decide.
Simulated run
- 14:32:19
The AP lead declines.
The supplier did not confirm the change, so the AP lead declines. The record is sealed and the example labelled.
- trail
- policy history
- examples
Simulated run
The cases are composites. The run is simulated.
FAQs
You might be wondering.
Questions
Answer
What is Gatehouse?
Gatehouse is software that sits between an AI agent and the systems it works in.
Before the agent acts, Gatehouse checks the action against rules your team has signed. Hard calls wait for a named person. Each action, hold and decision is saved in a sealed record.
Surehand, a small firm that deploys and runs AI agents for clients, built it and runs it.
Who is it for?
Teams that want AI agents to do real work, such as paying invoices or sorting claims, and need their security, compliance and finance reviewers to agree first.
If nobody will ever ask you to prove what an agent did, you probably do not need it.
How does the policy check work?
Your team writes one signed file for each deployment. We call it the manifest. It sets five things: scope, role, limits, review and seal.
Before each action runs, the policy engine checks it against that file. Allowed actions go ahead. Anything the file does not name is refused, and the refusal is saved to the record.
How the policy engine worksWhat happens when a rule holds an action?
The run pauses and the case goes to one named approver, not a shared inbox. The evidence is gathered before they are asked.
In our simulated run, a 2,400.00 invoice is held because the supplier's bank detail was first seen that day. The AP lead sees the invoice, the old detail and the new one, and declines the payment.
The decision and the reason are saved with the approver's name.
Follow the run step by stepCan someone override a block?
Yes, but not alone. Overriding a block takes two names.
Both names are saved with the override, next to the rule it overrode.
How do we know nobody changed the record?
Each record holds the SHA-256 digest of the one before it. If anyone edits a record later, us included, the chain breaks and the break shows.
Your team can check the chain on its own side, without Gatehouse installed.
Try to break the chainWhere does our data go?
Your business material stays in your own store. The audit trail holds links to it, not copies.
Material reaches a model provider only inside the scope the manifest allows, under API terms that exclude training. Each crossing is logged.
Is our data used to train models?
No. Under written API terms, ours and our providers', your data is not used to train any model.
Do you have SOC 2 or ISO 27001?
Not today. We have no SOC 2 audit report and no ISO 27001 certificate.
What we can offer now: a signed NDA and DPA before any business material moves, security questionnaires answered in writing with a name on them, and a record your team can check for itself.
When that changes, the security page will say so.
What is in place todayWhat do we keep if we stop using Gatehouse?
All of your records. The audit trail, policy history and labelled examples export in open formats.
You can read them without Gatehouse installed, so leaving does not mean losing your evidence.
Are the screens and cases on this site real?
The screens are simulated, and the run you see, RUN-7F2K4, is a worked example. Each one is labelled.
The three cases are composites of client work, with names and figures changed. No customer data appears anywhere on this site.
The three casesWhat happens in a walkthrough?
A 30-minute call with no slides. We bring manifest v14 and the sealed record of INV-40218. You bring one live process and the questions your reviewers will ask.
We usually reply to a request inside two business days.
Request a walkthroughHow is Gatehouse related to Surehand?
Surehand built Gatehouse, and each client deployment Surehand runs sits on it.
If you want an agent built and run for you, Surehand starts with a teardown of one process.
Visit Surehand